I just spent the past summer doing research at the 22nm level (designing L1/L2 caches with DVFS and other low-power techniques) and I can't agree more on SPICE/HSPICE's inability to converge.
I shrunk my designs down to the criitical paths (~12k transistors), and even providing the proper nodesets/initial conditions HSPICE was unable to converge or segfaulted quikcly. Fortunately, my university has a deal with Cadence through their University Alliance program -- Spectre may not be quite as accurate at HSPICE for analog circuits, but both it and Ultrasim (a FASTSPICE simulator for large designs) can handle much larger digital designs without complaint.
To the original submitter: Is there a good reason behind the no network connection requirement? If the university has a proper setup, students should be fine either on or off campus -- then it may be worth checking if your university has any deals with either Cadence, Synopsys, or Magma -- their tools are primarily Unix-based (Solaris, AIX, and Linux support), so it's just a matter of having the students SSH in with X forwarding or use VNC. This would even allows users with underpowered machines to simulate large designs quickly since everything is done remotely. I primarily run Windows on my local box, but either VMs with Linux or using Putty with Xming work properly for all these tools.